Add SFT pipeline: 302K clean dataset + auto-download pretrained from HF

LUNA 100M: cloud-ready training pipeline

Initial commit